Download i by Robert De Caux Supervised by Robin Hirsch September 2001

Transcript
Robert De Caux – MSc CS
Conclusions and Evaluation
8 Conclusions and Evaluation
8.1 Conclusions
The use of genetic programming allows optimal strategies to be generated very successfully
against a selection of opponents. The random principles of evolution mean that it is
sometimes difficult to evolve the more complex strategies, but by adjusting the GP parameters
to increase diversity, they are usually discovered. The fact that the whole game history is
available is used by some of the more complex strategies, and this is where GP has the
advantage over GA.
The experiments with coevolution show that without the If, EQ and Go functions, cooperative
behaviour emerges as the equilibrium from a random selection of Individuals. This is because
although defective Individuals can gain short-term benefits by exploitation of cooperative
players, they struggle over the long term as they perform badly against players similar to
themselves. On the other hand, cooperative players perform well amongst themselves, and so
do well over the long term. The difficulty is in establishing the cooperative behaviour in the
first place, as this requires enough non-defecting strategies to establish a foothold. These
strategies must be robust enough to perform well against defective players to have a chance of
selection, but then cooperate with each other. This is what makes Tit-For-Tat so successful.
Once in a cooperative situation, any cooperating player will score well, but unless enough of
the Population are robust enough to prevent invasion, defective strategies can alter this
equilibrium.
When Individuals are allowed to take into account which go of the game they are on,
opportunistic strategies are able to disrupt the cooperative equilibrium if they defect on the
final go when retribution is not possible. This can allow defective players to take over, but
usually more robust versions of the (Go Last) strategy take over and form a quasi-equilbrium
around 3.0. This still means that there is a very large amount of cooperation in the
Population.
Allowing Individuals to hunt for opponents tends to encourage cooperation, as cooperative
players make the best opponents. Once established, this cooperative equilibrium is much
harder to disrupt, as defecting players cannot find opponents to play and so score badly.
There is a very delicate threshold for setting the playability however, above which finding an
opponent becomes the most important factor.
58